Automated creation of pattern database search heuristics

Abstract

Pattern databases are dictionaries for heuristic estimates storing state-to-goal distances in state space abstractions. Their effectiveness is sensitive to the selection of the underlying patterns. Especially for multiple and additive pattern databases, the manual selection of patterns that leads to good exploration results is involved. For automating the selection process, greedy bin-packing has been suggested. This paper proposes genetic algorithms to optimize its output. Patterns are encoded as binary strings and optimized using an objective function that predicts the heuristic search tree size based on the distribution of heuristic values in abstract space. To reduce the memory requirements we construct the pattern databases symbolically. Experiments in heuristic search planning indicate that the total search efforts can be reduced significantly. © Springer-Verlag Berlin Heidelberg 2007.